urcu-qsbr.c

   1 /*
   2  * urcu-qsbr.c
   3  *
   4  * Userspace RCU QSBR library
   5  *
   6  * Copyright (c) 2009 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
   7  * Copyright (c) 2009 Paul E. McKenney, IBM Corporation.
   8  *
   9  * This library is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public
  11  * License as published by the Free Software Foundation; either
  12  * version 2.1 of the License, or (at your option) any later version.
  13  *
  14  * This library is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with this library; if not, write to the Free Software
  21  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  22  *
  23  * IBM's contributions to this file may be relicensed under LGPLv2 or later.
  24  */
  25
  26 #define _GNU_SOURCE
  27 #define _LGPL_SOURCE
  28 #include <stdio.h>
  29 #include <pthread.h>
  30 #include <signal.h>
  31 #include <assert.h>
  32 #include <stdlib.h>
  33 #include <stdint.h>
  34 #include <string.h>
  35 #include <errno.h>
  36 #include <poll.h>
  37
  38 #include "urcu/wfqueue.h"
  39 #include "urcu/map/urcu-qsbr.h"
  40 #define BUILD_QSBR_LIB
  41 #include "urcu/static/urcu-qsbr.h"
  42 #include "urcu-pointer.h"
  43 #include "urcu/tls-compat.h"
  44
  45 #include "urcu-die.h"
  46
  47 /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
  48 #undef _LGPL_SOURCE
  49 #include "urcu-qsbr.h"
  50 #define _LGPL_SOURCE
  51
  52 void __attribute__((destructor)) rcu_exit(void);
  53
  54 /*
  55  * rcu_gp_lock ensures mutual exclusion between threads calling
  56  * synchronize_rcu().
  57  */
  58 static pthread_mutex_t rcu_gp_lock = PTHREAD_MUTEX_INITIALIZER;
  59 /*
  60  * rcu_registry_lock ensures mutual exclusion between threads
  61  * registering and unregistering themselves to/from the registry, and
  62  * with threads reading that registry from synchronize_rcu(). However,
  63  * this lock is not held all the way through the completion of awaiting
  64  * for the grace period. It is sporadically released between iterations
  65  * on the registry.
  66  * rcu_registry_lock may nest inside rcu_gp_lock.
  67  */
  68 static pthread_mutex_t rcu_registry_lock = PTHREAD_MUTEX_INITIALIZER;
  69
  70 int32_t gp_futex;
  71
  72 /*
  73  * Global grace period counter.
  74  */
  75 unsigned long rcu_gp_ctr = RCU_GP_ONLINE;
  76
  77 /*
  78  * Active attempts to check for reader Q.S. before calling futex().
  79  */
  80 #define RCU_QS_ACTIVE_ATTEMPTS 100
  81
  82 /*
  83  * Written to only by each individual reader. Read by both the reader and the
  84  * writers.
  85  */
  86 __DEFINE_URCU_TLS_GLOBAL(struct rcu_reader, rcu_reader);
  87
  88 #ifdef DEBUG_YIELD
  89 unsigned int yield_active;
  90 __DEFINE_URCU_TLS_GLOBAL(unsigned int, rand_yield);
  91 #endif
  92
  93 static CDS_LIST_HEAD(registry);
  94
  95 static void mutex_lock(pthread_mutex_t *mutex)
  96 {
  97         int ret;
  98
  99 #ifndef DISTRUST_SIGNALS_EXTREME
 100         ret = pthread_mutex_lock(mutex);
 101         if (ret)
 102                 urcu_die(ret);
 103 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
 104         while ((ret = pthread_mutex_trylock(mutex)) != 0) {
 105                 if (ret != EBUSY && ret != EINTR)
 106                         urcu_die(ret);
 107                 poll(NULL,0,10);
 108         }
 109 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
 110 }
 111
 112 static void mutex_unlock(pthread_mutex_t *mutex)
 113 {
 114         int ret;
 115
 116         ret = pthread_mutex_unlock(mutex);
 117         if (ret)
 118                 urcu_die(ret);
 119 }
 120
 121 /*
 122  * synchronize_rcu() waiting. Single thread.
 123  */
 124 static void wait_gp(void)
 125 {
 126         /* Read reader_gp before read futex */
 127         cmm_smp_rmb();
 128         if (uatomic_read(&gp_futex) == -1)
 129                 futex_noasync(&gp_futex, FUTEX_WAIT, -1,
 130                       NULL, NULL, 0);
 131 }
 132
 133 /*
 134  * Always called with rcu_registry lock held. Releases this lock between
 135  * iterations and grabs it again. Holds the lock when it returns.
 136  */
 137 static void update_counter_and_wait(void)
 138 {
 139         CDS_LIST_HEAD(qsreaders);
 140         unsigned int wait_loops = 0;
 141         struct rcu_reader *index, *tmp;
 142
 143 #if (CAA_BITS_PER_LONG < 64)
 144         /* Switch parity: 0 -> 1, 1 -> 0 */
 145         CMM_STORE_SHARED(rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR);
 146 #else   /* !(CAA_BITS_PER_LONG < 64) */
 147         /* Increment current G.P. */
 148         CMM_STORE_SHARED(rcu_gp_ctr, rcu_gp_ctr + RCU_GP_CTR);
 149 #endif  /* !(CAA_BITS_PER_LONG < 64) */
 150
 151         /*
 152          * Must commit rcu_gp_ctr update to memory before waiting for
 153          * quiescent state. Failure to do so could result in the writer
 154          * waiting forever while new readers are always accessing data
 155          * (no progress). Enforce compiler-order of store to rcu_gp_ctr
 156          * before load URCU_TLS(rcu_reader).ctr.
 157          */
 158         cmm_barrier();
 159
 160         /*
 161          * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
 162          * model easier to understand. It does not have a big performance impact
 163          * anyway, given this is the write-side.
 164          */
 165         cmm_smp_mb();
 166
 167         /*
 168          * Wait for each thread rcu_reader_qs_gp count to become 0.
 169          */
 170         for (;;) {
 171                 if (wait_loops < RCU_QS_ACTIVE_ATTEMPTS)
 172                         wait_loops++;
 173                 if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
 174                         uatomic_set(&gp_futex, -1);
 175                         /*
 176                          * Write futex before write waiting (the other side
 177                          * reads them in the opposite order).
 178                          */
 179                         cmm_smp_wmb();
 180                         cds_list_for_each_entry(index, &registry, node) {
 181                                 _CMM_STORE_SHARED(index->waiting, 1);
 182                         }
 183                         /* Write futex before read reader_gp */
 184                         cmm_smp_mb();
 185                 }
 186                 cds_list_for_each_entry_safe(index, tmp, &registry, node) {
 187                         if (!rcu_gp_ongoing(&index->ctr))
 188                                 cds_list_move(&index->node, &qsreaders);
 189                 }
 190
 191                 if (cds_list_empty(&registry)) {
 192                         if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
 193                                 /* Read reader_gp before write futex */
 194                                 cmm_smp_mb();
 195                                 uatomic_set(&gp_futex, 0);
 196                         }
 197                         break;
 198                 } else {
 199                         /* Temporarily unlock the registry lock. */
 200                         mutex_unlock(&rcu_registry_lock);
 201                         if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
 202                                 wait_gp();
 203                         } else {
 204 #ifndef HAS_INCOHERENT_CACHES
 205                                 caa_cpu_relax();
 206 #else /* #ifndef HAS_INCOHERENT_CACHES */
 207                                 cmm_smp_mb();
 208 #endif /* #else #ifndef HAS_INCOHERENT_CACHES */
 209                         }
 210                         /* Re-lock the registry lock before the next loop. */
 211                         mutex_lock(&rcu_registry_lock);
 212                 }
 213         }
 214         /* put back the reader list in the registry */
 215         cds_list_splice(&qsreaders, &registry);
 216 }
 217
 218 /*
 219  * Using a two-subphases algorithm for architectures with smaller than 64-bit
 220  * long-size to ensure we do not encounter an overflow bug.
 221  */
 222
 223 #if (CAA_BITS_PER_LONG < 64)
 224 void synchronize_rcu(void)
 225 {
 226         unsigned long was_online;
 227
 228         was_online = URCU_TLS(rcu_reader).ctr;
 229
 230         /* All threads should read qparity before accessing data structure
 231          * where new ptr points to.  In the "then" case, rcu_thread_offline
 232          * includes a memory barrier.
 233          *
 234          * Mark the writer thread offline to make sure we don't wait for
 235          * our own quiescent state. This allows using synchronize_rcu()
 236          * in threads registered as readers.
 237          */
 238         if (was_online)
 239                 rcu_thread_offline();
 240         else
 241                 cmm_smp_mb();
 242
 243         mutex_lock(&rcu_gp_lock);
 244         mutex_lock(&rcu_registry_lock);
 245
 246         if (cds_list_empty(&registry))
 247                 goto out;
 248
 249         /*
 250          * Wait for previous parity to be empty of readers.
 251          * update_counter_and_wait() can release and grab again
 252          * rcu_registry_lock interally.
 253          */
 254         update_counter_and_wait();      /* 0 -> 1, wait readers in parity 0 */
 255
 256         /*
 257          * Must finish waiting for quiescent state for parity 0 before
 258          * committing next rcu_gp_ctr update to memory. Failure to
 259          * do so could result in the writer waiting forever while new
 260          * readers are always accessing data (no progress).  Enforce
 261          * compiler-order of load URCU_TLS(rcu_reader).ctr before store to
 262          * rcu_gp_ctr.
 263          */
 264         cmm_barrier();
 265
 266         /*
 267          * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
 268          * model easier to understand. It does not have a big performance impact
 269          * anyway, given this is the write-side.
 270          */
 271         cmm_smp_mb();
 272
 273         /*
 274          * Wait for previous parity to be empty of readers.
 275          * update_counter_and_wait() can release and grab again
 276          * rcu_registry_lock interally.
 277          */
 278         update_counter_and_wait();      /* 1 -> 0, wait readers in parity 1 */
 279 out:
 280         mutex_unlock(&rcu_registry_lock);
 281         mutex_unlock(&rcu_gp_lock);
 282
 283         /*
 284          * Finish waiting for reader threads before letting the old ptr being
 285          * freed.
 286          */
 287         if (was_online)
 288                 rcu_thread_online();
 289         else
 290                 cmm_smp_mb();
 291 }
 292 #else /* !(CAA_BITS_PER_LONG < 64) */
 293 void synchronize_rcu(void)
 294 {
 295         unsigned long was_online;
 296
 297         was_online = URCU_TLS(rcu_reader).ctr;
 298
 299         /*
 300          * Mark the writer thread offline to make sure we don't wait for
 301          * our own quiescent state. This allows using synchronize_rcu()
 302          * in threads registered as readers.
 303          */
 304         if (was_online)
 305                 rcu_thread_offline();
 306         else
 307                 cmm_smp_mb();
 308
 309         mutex_lock(&rcu_gp_lock);
 310         mutex_lock(&rcu_registry_lock);
 311         if (cds_list_empty(&registry))
 312                 goto out;
 313         /*
 314          * update_counter_and_wait() can release and grab again
 315          * rcu_registry_lock interally.
 316          */
 317         update_counter_and_wait();
 318 out:
 319         mutex_unlock(&rcu_registry_lock);
 320         mutex_unlock(&rcu_gp_lock);
 321
 322         if (was_online)
 323                 rcu_thread_online();
 324         else
 325                 cmm_smp_mb();
 326 }
 327 #endif  /* !(CAA_BITS_PER_LONG < 64) */
 328
 329 /*
 330  * library wrappers to be used by non-LGPL compatible source code.
 331  */
 332
 333 void rcu_read_lock(void)
 334 {
 335         _rcu_read_lock();
 336 }
 337
 338 void rcu_read_unlock(void)
 339 {
 340         _rcu_read_unlock();
 341 }
 342
 343 void rcu_quiescent_state(void)
 344 {
 345         _rcu_quiescent_state();
 346 }
 347
 348 void rcu_thread_offline(void)
 349 {
 350         _rcu_thread_offline();
 351 }
 352
 353 void rcu_thread_online(void)
 354 {
 355         _rcu_thread_online();
 356 }
 357
 358 void rcu_register_thread(void)
 359 {
 360         URCU_TLS(rcu_reader).tid = pthread_self();
 361         assert(URCU_TLS(rcu_reader).ctr == 0);
 362
 363         mutex_lock(&rcu_registry_lock);
 364         cds_list_add(&URCU_TLS(rcu_reader).node, &registry);
 365         mutex_unlock(&rcu_registry_lock);
 366         _rcu_thread_online();
 367 }
 368
 369 void rcu_unregister_thread(void)
 370 {
 371         /*
 372          * We have to make the thread offline otherwise we end up dealocking
 373          * with a waiting writer.
 374          */
 375         _rcu_thread_offline();
 376         mutex_lock(&rcu_registry_lock);
 377         cds_list_del(&URCU_TLS(rcu_reader).node);
 378         mutex_unlock(&rcu_registry_lock);
 379 }
 380
 381 void rcu_exit(void)
 382 {
 383         /*
 384          * Assertion disabled because call_rcu threads are now rcu
 385          * readers, and left running at exit.
 386          * assert(cds_list_empty(&registry));
 387          */
 388 }
 389
 390 DEFINE_RCU_FLAVOR(rcu_flavor);
 391
 392 #include "urcu-call-rcu-impl.h"
 393 #include "urcu-defer-impl.h"